1 00:00:00,000 --> 00:00:03,520 Right, so this one's brought to you by Jellypod AI, and it's a follow up on 2 00:00:03,560 --> 00:00:05,440 something we skated past. 3 00:00:05,488 --> 00:00:10,160 Last time, in the subagent bloat episode, we tossed allow_large in as a closing 4 00:00:10,217 --> 00:00:10,800 bullet. 5 00:00:10,880 --> 00:00:13,120 Turns out there's more to it than a bullet. 6 00:00:13,125 --> 00:00:13,445 Yeah. 7 00:00:13,468 --> 00:00:20,245 Claude Code 2.1.296 shipped October 9th, and the release note says Claude can read a 8 00:00:20,309 --> 00:00:24,805 text file past the usual size limits in one call, when it needs the whole file and 9 00:00:24,825 --> 00:00:25,925 the context has room. 10 00:00:26,072 --> 00:00:30,405 Classmethod then actually tested it, and the numbers are the interesting part. 11 00:00:30,417 --> 00:00:31,377 Start with the problem. 12 00:00:31,537 --> 00:00:34,017 Read has a token ceiling per call. 13 00:00:34,204 --> 00:00:38,817 Point it at a big log or a generated fixture and you get a slice back, 14 00:00:38,897 --> 00:00:42,257 with a flag saying truncatedByTokenCap is true. 15 00:00:42,470 --> 00:00:48,177 Classmethod made a 3,000 line text file, about 204 kilobytes, 16 00:00:48,217 --> 00:00:50,897 and Read stopped at line 732. 17 00:00:51,117 --> 00:00:57,137 The file shows as 3,001 total lines because Read counts the empty bit after the 18 00:00:57,204 --> 00:00:58,097 final newline. 19 00:00:58,125 --> 00:01:00,605 And a truncated read is a sneaky failure. 20 00:01:00,731 --> 00:01:05,405 Claude either works from a partial view, or it burns turns paging through with 21 00:01:05,485 --> 00:01:06,805 offset and limit. 22 00:01:06,877 --> 00:01:12,845 With allow_large, in their second run, Claude read all 3,001 lines in one call, 23 00:01:12,865 --> 00:01:14,925 and the truncation flag came back null. 24 00:01:14,958 --> 00:01:16,478 Here's the bit that surprised me. 25 00:01:16,718 --> 00:01:17,998 It was cheaper. 26 00:01:18,265 --> 00:01:23,518 On Opus 5.5 the truncated run cost about six and a half cents, 27 00:01:23,538 --> 00:01:26,318 and the full read cost about five and a bit. 28 00:01:26,430 --> 00:01:29,358 That's 19.1 percent lower. 29 00:01:29,598 --> 00:01:35,998 On Haiku 5.5 it went from about 0.19 cents to 0.14 cents, 30 00:01:36,318 --> 00:01:38,398 23.8 percent lower. 31 00:01:38,557 --> 00:01:40,157 Now, caution on interpretation. 32 00:01:40,817 --> 00:01:43,457 Classmethod reported the numbers, they didn't prove why. 33 00:01:44,177 --> 00:01:48,257 My read is that a complete first read means fewer clarifying moves afterward, 34 00:01:48,697 --> 00:01:50,777 but that's a guess, not something the test showed. 35 00:01:51,797 --> 00:01:55,337 Also, that was one small synthetic file with max turns set to one. 36 00:01:55,977 --> 00:01:56,917 Don't take it as a law. 37 00:01:57,625 --> 00:01:58,025 Fair. 38 00:01:58,125 --> 00:01:59,865 The model gap was steady though. 39 00:02:00,025 --> 00:02:05,625 Opus ran roughly 34 to 37 times the cost of Haiku on the same job. 40 00:02:05,625 --> 00:02:06,425 So how do you use it? 41 00:02:06,558 --> 00:02:10,265 First thing: allow_large isn't a CLI flag. 42 00:02:10,356 --> 00:02:13,065 Claude sets it as a parameter when it calls Read. 43 00:02:13,185 --> 00:02:16,105 You trigger it by saying plainly that you need the whole file. 44 00:02:16,238 --> 00:02:21,065 Classmethod's prompt was: the entire big.txt file is needed, 45 00:02:21,111 --> 00:02:25,065 please read the entire file in a single Read call without omitting anything. 46 00:02:25,165 --> 00:02:29,465 Without mentioning allow_large at all, Claude set it to true on its own. 47 00:02:29,500 --> 00:02:33,180 And you can bake that into CLAUDE.md or a subagent's instructions. 48 00:02:33,252 --> 00:02:38,780 Something like: when inspecting build logs, schema definitions or test fixtures, 49 00:02:38,908 --> 00:02:41,740 read the complete file in a single Read call. 50 00:02:41,884 --> 00:02:44,540 Then you're not retyping it every session. 51 00:02:44,542 --> 00:02:47,102 If you want proof it fired, run headless. 52 00:02:47,222 --> 00:02:51,902 Use claude with the p flag, max turns one, allowedTools Read, 53 00:02:51,959 --> 00:02:56,502 stream JSON output format with verbose on, and send your prompt. 54 00:02:56,551 --> 00:03:01,022 Redirect to a file, then use jq to pull the tool_use entries. 55 00:03:01,122 --> 00:03:04,462 You're looking for tool Read, allow_large true. 56 00:03:04,595 --> 00:03:08,542 In the tool result, check that truncatedByTokenCap is null. 57 00:03:08,542 --> 00:03:09,822 Quick heads up from their run. 58 00:03:09,922 --> 00:03:13,982 The exit code was 1, but that's just because max turns one stopped it. 59 00:03:14,082 --> 00:03:15,182 Not a failure. 60 00:03:15,208 --> 00:03:16,088 Now the caveats. 61 00:03:16,228 --> 00:03:20,968 The changelog says when the context has room, and that's a real condition. 62 00:03:21,128 --> 00:03:25,848 If you're deep in a session near compaction, a big read has less space to land. 63 00:03:25,988 --> 00:03:30,808 And everything you read in stays in context across turns, so a huge file keeps 64 00:03:30,868 --> 00:03:31,608 costing you. 65 00:03:31,821 --> 00:03:34,169 It's also text files only. 66 00:03:34,208 --> 00:03:36,448 So use it when you truly need the whole thing. 67 00:03:36,581 --> 00:03:40,528 If you only need one stack trace, grep it, don't swallow the lot. 68 00:03:40,542 --> 00:03:43,982 Few other things in 2.1.296 worth a mention. 69 00:03:44,142 --> 00:03:49,342 Edit and NotebookEdit now refuse files that aren't valid UTF 8, 70 00:03:49,374 --> 00:03:53,262 like Windows 1252, Shift JIS and GBK. 71 00:03:53,411 --> 00:03:57,662 Before, they replaced every non ASCII character in those files. 72 00:03:57,822 --> 00:04:01,342 Now the edit's rejected, so check your legacy encodings. 73 00:04:01,375 --> 00:04:06,015 On Windows, PowerShell commands over about one kilobyte used to prompt every time. 74 00:04:06,162 --> 00:04:11,295 Allow rules and read only detection now apply up to 32 kilobytes. 75 00:04:11,482 --> 00:04:18,415 There's a security fix too: Bash commands that assign to BASH_ARGV0 and then use it 76 00:04:18,527 --> 00:04:20,255 were being auto approved. 77 00:04:20,415 --> 00:04:22,015 They prompt now. 78 00:04:23,000 --> 00:04:27,000 And with the debug flag, hooks now log their command, plugin, 79 00:04:27,050 --> 00:04:29,080 outcome and duration when they finish. 80 00:04:29,267 --> 00:04:32,760 So if a custom hook is dragging, you can finally see which one. 81 00:04:32,874 --> 00:04:36,440 Update, then try the whole file prompt on your biggest log.